Imagination is a part of daily life. We use it successfully every day to avoid disaster, even with very minor actions. Imagining what can go wrong, we choose safer options. Of course some of us are better at this than others, but you get the point, don’t you? Imagination is essential to rationality.

DeepMind is talking about designing the ability to imagine into its artificial intelligence for this very reason. The ability to predict the probable outcomes of each of a series of courses of action, then choose the one with the most desirable outcome could make AI much more powerful and safer at the same time. But this assumes that the criteria for choosing the most desirable outcome are the right ones, an assumption that needs careful analysis.

You can read about what DeepMind is doing in the article, “Agents that imagine and plan” on DeepMind’s own website.